Auteur : Luca Saiu (LIPN - P13)

Titre : 
Scalable BIBOP garbage collection for parallel functional programs on multi-core machines

Résumé : 

Exploiting modern multi-core processors requires task-parallel
programs which are simply too hard to implement with traditional
techniques; as high-level languages rely on automatic memory
management subsystems, such subsystems must be made fast and scalable,
tuned for today's complex memory architectures.  I describe my
implementation of a new parallel non-moving garbage collector for
shared-memory machines, based on a variant of the BIBOP
organization. Building on the experience of Boehm's work and
revisiting some older ideas in the light of current hardware
performance models, I propose several algorithmic improvements leading
to compact data representation and measurable speedups, particularly
in the context of functional programs.  While discussing in detail the
performance-critical sections of the implementation I provide an
intuitive justification for my choices which I then corroborate with
measurements: allocation-heavy benchmarks average a scalability of
7.41 on eight cores.  This effort results in a particularly clean
architecture based on just a few data structures, which lends itself
to experimentation with alternative techniques.